Back to EveryPatent.com
United States Patent |
6,216,083
|
Ulyanov
,   et al.
|
April 10, 2001
|
System for intelligent control of an engine based on soft computing
Abstract
A reduced control system suitable for control of an engine as a nonlinear
plant is described. The reduced control system is configured to use a
reduced sensor set for controlling the plant without significant loss of
control quality (accuracy) as compared to an optimal control system with
an optimum sensor set. The control system calculates the information
content provided by the reduced sensor set as compared to the information
content provided by the optimum set. The control system also calculates
the difference between the entropy production rate of the plant and the
entropy production rate of the controller. A genetic optimizer is used to
tune a fuzzy neural network in the reduced controller. A fitness function
for the genetic optimizer provides optimum control accuracy in the reduced
control system by minimizing the difference in entropy production while
maximizing the sensor information content.
Inventors:
|
Ulyanov; Sergei V. (Shizuoka, JP);
Hashimoto; Shigeki (Shizuoka, JP);
Yamaguchi; Masashi (Shizuoka, JP)
|
Assignee:
|
Yamaha Motor Co., Ltd. (Shizuoka, JP)
|
Appl. No.:
|
176987 |
Filed:
|
October 22, 1998 |
Current U.S. Class: |
701/106; 701/115 |
Intern'l Class: |
F02D 041/14; G06F 019/00 |
Field of Search: |
701/106,109,115
123/674,675
|
References Cited
U.S. Patent Documents
5361628 | Nov., 1994 | Marko et al. | 73/116.
|
5539638 | Jul., 1996 | Keeler et al. | 123/672.
|
Primary Examiner: Dolinar; Andrew M.
Attorney, Agent or Firm: Knobbe, Martens, Olson & Bear, LLP.
Claims
What is claimed is:
1. A method for controlling an internal combustion engine comprising the
steps of: measuring first information from said engine by using a first
plurality of sensors; providing said first information to a first engine
control system, said first engine control system configured to provide a
desired accuracy for said engine, said first control system providing a
first control signal; measuring second information from said engine by
using a second plurality of sensors, where said second plurality of
sensors comprises fewer sensors that said first plurality of sensors,
providing said second information to a second engine control systems said
second engine control system providing a second control signal; and
configuring said second engine control system using said first engine
controls signal and said second engine control signal by generating a
physical criteria and an information criteria, wherein said physical
criteria is calculated by an entropy model based on thermodynamic
properties of said engine.
2. The method of claim 1, wherein said physical criteria is calculated by
an entropy model that calculates an entropy production rate.
3. The method of claim 2, wherein said second control system is adapted to
reduce an entropy production in said second control system and said
engine.
4. The method of claim 3, wherein said thermodynamic model is based on
engine air temperature and engine water temperature.
5. A method for controlling an internal combustion engine, comprising the
steps of: measuring first information from said engine by using a first
plurality of sensors; providing said first information to a first engine
control system said first engine control system configured to provide a
desired accuracy for said engine, said first control system providing a
first control signal; measuring second information from said engine by
using a second plurality of sensors, where said second plurality of
sensors comprises fewer sensors that said first plurality of sensors,
providing said second information to a second engine control system, said
second engine control system providing a second control signal; and
confining said second engine control system using said first engine
controls signal and said second engine control signal by generating a
physical criteria and an information criteria, wherein said optimizer uses
a genetic algorithm having a fitness function, wherein a portion of said
fitness function is based on entropy.
6. A method for controlling an internal combustion engine, comprising the
steps of: measuring first information from said engine by using a first
plurality of sensors; providing said first information to a first engine
control system, said first engine control system configured to provide a
desired accuracy for said engine, said first control system providing a
first control signal; measuring second information from said engine by
using a second plurality of sensors, where said second plurality of
sensors comprises fewer sensors that said first plurality of sensors,
providing said second information to a second engine control system, said
second engine control system providing a second control signal; and
configuring said second engine control system using said first engine
controls signal and said second engine control signal be generating a
physical criteria and an information criteria, wherein said step of
configuring further comprises providing said physical criteria and said
information criteria to a genetic algorithm having a fitness function,
said fitness function based on entropy.
7. A control system adapted to control an engine, comprising: a first
plurality of sensors configured to measure first information about said
engine, a first engine controller configured to receive at least a portion
of said first information, said first engine controller trained to produce
a first control signal, where said first engine controller is trained to
use relatively more of said at least a portion of said first information
signal in order to reduce an entropy production of said engine, wherein
said first engine controller comprises a fuzzy neural network configured
to be trained by a genetic analyzer having a first fitness function, said
first fitness function configured to increase mutual information between
said first control signal and a second control signal, said second control
signal provided by a second controller configured to receive information
from a second plurality of sensors, wherein said second plurality of
sensors is greater than said first plurality of sensors.
8. The apparatus of claim 7, wherein said genetic analyzer further
comprises a second fitness function configured to reduce entropy
production rate of said engine.
9. The apparatus of claim 8, wherein said genetic analyzer is configured to
use said second fitness function to realize a node correction in said
fuzzy neural network.
10. The apparatus of claim 7, wherein said first plurality of sensors
comprises a temperature sensor.
11. The apparatus of claim 7, wherein said second plurality of sensors
comprises an oxygen sensor.
12. The apparatus of claim 7, wherein said first plurality of sensors
comprises a water temperature sensor, an air temperature sensor, and an
airflow sensor.
13. The apparatus of claim 7, wherein said first plurality of sensors
comprises a water temperature sensor, an air temperature sensor, an
airflow sensor, and an oxygen sensor.
14. The apparatus of claim 7, wherein said first control signal comprises
an injector control signal configured to control a fuel injector.
Description
BACKGROUND OF THE INVENTION
1. Field of the Invention
The disclosed invention relates generally to engine control systems, and
more particularly to electronic control systems for internal combustion
engines.
2. Description of the Related Art
Feedback control systems are widely used to maintain the output of a
dynamic system at a desired value in spite of external disturbance forces
that would move the output away from the desired value. For example, a
household furnace controlled by a thermostat is an example of a, feedback
control system. The thermostat continuously measures the air temperature
of the house, and when the temperature falls below a desired minimum
temperature, the thermostat turns the furnace on. When the furnace has
warmed the air above the desired minimum temperature, then the thermostat
turns the furnace off. The thermostat-furnace system maintains the
household temperature at a constant value in spite of external
disturbances such as a drop in the outside air temperature. Similar types
of feedback control are used in many applications.
A central component in a feedback control system is a controlled object,
otherwise known as a process "plant," whose output variable is to be
controlled. In the above example, the plant is; the house, the output
variable is the air temperature of the house, and the disturbance is the
flow of heat through the walls of the house. The plant is controlled by a
control system. In the above example, the control system is the thermostat
in combination with the furnace. The thermostat-furnace system uses simple
on-off feedback control to maintain the temperature of the house. In many
control environments, such as motor shaft position or motor speed control
systems, simple on-off feedback control is insufficient. More advanced
control systems rely on combinations of proportional feedback control,
integral feedback control, and derivative feedback control.
The PID control system is a linear control system that is based on a
dynamic model of the plant. In classical control systems, a linear dynamic
model is obtained in the form of dynamic equations, usually ordinary
differential equations. The plant is assumed to be relatively linear, time
invariant, and stable. However, many real-world plants are time varying,
highly nonlinear, and unstable. For example, the dynamic model may contain
parameters (e.g., masses, inductances, aerodynamic coefficients, etc.)
which are either poorly known or depend on a changing environment. If the
parameter variation is small and the dynamic model is stable, then the PID
controller may be sufficient. However, if the parameter variation is
large, or if the dynamic model is unstable, then it is common to add
adaptation or intelligent (AI) control to the PID control system.
AI control systems use an optimizer, typically a nonlinear optimizer, to
program the operation of the PID controller and thereby improve the
overall operation of the control system. The optimizers used in many AI
control systems rely on a genetic algorithm. Using a set of inputs, and a
fitness function, the genetic algorithm works in a manner similar to
process of evolution to arrive at a solution which is, hopefully, optimal.
The genetic algorithm generates sets of chromosomes (corresponding to
possible solutions) and then sorts the chromosomes by evaluating each
solution using the fitness function. The fitness function determines where
each solution ranks on a fitness scale. Chromosomes which are more fit,
are those chromosomes which correspond to solutions that rate high on the
fitness scale. Chromosomes which are less fit, are those chromosomes which
correspond to solutions that rate low on the fitness scale. Chromosomes
that are more fit are kept (survive) and chromosomes that are less fit are
discarded (die). New chromosomes are created to replace the discarded
chromosomes. The new chromosomes are created by crossing pieces of
existing chromosomes and by introducing mutations.
The PID controller has a linear transfer function and thus is based upon a
linearized equation of motion for the plant. Prior art genetic algorithms
used to program PID controllers typically use simple fitness functions and
thus do not solve the problem of poor controllability typically seen in
linearization models. As is the case with most optimizers, the success or
failure of the optimization often ultimately depends on the selection of
the performance (fitness) function.
Evaluating the motion characteristics of a nonlinear plant is often
difficult, in part due to the lack of a general analysis method.
Conventionally, when controlling a plant with nonlinear motion
characteristics, it is common to find certain equilibrium points of the
plant and the motion characteristics of the plant are linearized in a
vicinity near an equilibrium point. Control is then based on evaluating
the pseudo (linearized) motion characteristics near the equilibrium point.
This. technique works poorly, if at all, for plants described by models
that are unstable or dissipative.
SUMMARY OF THE INVENTION
The present invention solves these and other problems by providing a new AI
control system that allows a reduced number of sensors to be used without
a significant loss in control accuracy. The new AI control system is
self-organizing and uses a fitness (performance) function that are based
on the physical laws of minimum entropy and maximum sensor information.
The self-organizing control system may be used to control complex plants
described by nonlinear, unstable, dissipative models. The reduced control
system is configured to use smart simulation techniques for controlling
the plant despite the reduction in the number of sensor number without
significant loss of control quality (accuracy) as compared to an optimal
control system. In one embodiment, the reduced control system comprises a
neural network that is trained by a genetic analyzer. The genetic analyzer
uses a fitness function that maximizes information while minimizing
entropy production.
In one embodiment, the reduced control system is applied to an internal
combustion engine to provide control without the use of extra sensors,
such as, for example, an oxygen sensor. The reduced control system
develops a reduced control signal from a reduced sensor set. The reduced
control system is trained by a genetic analyzer that uses a control signal
developed by an optimized control system. The optimized control system
provides an optimum control signal based on data obtained from temperature
sensors, air-flow sensors, and an oxygen sensor. In an off-line learning
mode, the optimum control signal is subtracted from a reduced control
signal (developed by the reduced control system) and provided to an
information calculator. The information calculator provides an information
criteria to the genetic analyzer. Data from the reduced sensor set is also
provided to an entropy model, which calculates a physical criteria based
on entropy. The physical criteria is also provided to the genetic
analyzer. The genetic analyzer uses both the information criteria and the
physical criteria to develop a training signal for the reduced control
system.
In one embodiment, a reduced control system is applied to a vehicle
suspension to provide control of the suspension system using data from a
reduced number of sensors. The reduced control system develops a reduced
control signal from a reduced sensor set. The reduced control system is
trained by a genetic analyzer that uses a control signal developed by an
optimized control system. The optimized control system provides an optimum
control signal based on data obtained from a plurality of angle and
position sensors. In an off-line learning mode, the optimum control signal
is subtracted from a reduced control signal (developed by the reduced
control system) and provided to an information calculator. In one
embodiment, the reduced control system uses a vertical accelerometer
mounted near the center of the vehicle. The information calculator
provides an information criteria to the genetic analyzer. Data from the
reduced sensor set is also provided to an entropy model, which calculates
a physical criteria based on entropy. The physical criteria is also
provided to the genetic analyzer. The genetic analyzer uses both the
information criteria and the physical criteria to develop a training
signal for the reduced control system.
In one embodiment, the invention includes a method for controlling a
nonlinear object (a plant) by obtaining an entropy production difference
between a time differentiation (dS.sub.u /dt) of the entropy of the plant
and a time differentiation (dS.sub.c /dt) of the entropy provided to the
plant from a controller. A genetic algorithm that uses the entropy
production difference as a fitness (performance) function evolves a
control rule for a low-level controller, such as a PID controller. The
nonlinear stability characteristics of the plant are evaluated using a
Lyapunov function. The evolved control rule may be corrected using further
evolutions using an information function that compares the information
available from an optimum sensor system with the information available
from a reduced sensor system. The genetic analyzer minimizes entropy and
maximizes sensor information content.
In some embodiments, the control method may also include evolving a control
rule relative to a variable of the controller by means of a genetic
algorithm. The genetic algorithm uses a fitness function based on a
difference between a time differentiation of the entropy of the plant
(dS.sub.u /dt) and a time differentiation (dS.sub.c /dt) of the entropy
provided to the plant. The variable may be corrected by using the evolved
control rule.
In another embodiment, the invention comprises an AI control system adapted
to control a nonlinear plant. The AI control system includes a simulator
configured to use a thermodynamic model of a nonlinear equation of motion
for the plant. The thermodynamic model is based on EL Lyapunov function
(V), and the simulator uses the function V to analyze control for a state
stability of the plant. The AI control system calculates an entropy
production difference between a time differentiation of the entropy of
said plant (dS.sub.u /dt) and a time differentiation (dS.sub.c /dt) of the
entropy provided to the plant by a low-level controller that controls the
plant. The entropy production difference is used by a genetic algorithm to
obtain an adaptation function in which the entropy production difference
is minimized. The genetic algorithm provides a teaching signal to a fuzzy
logic classifier that determines a fuzzy rule by using a learning process.
The fry logic controller is also configured to form a control rule that
sets a control variable of the low-level controller.
In one embodiment, the low-level controller is a linear controller such as
a PID controller. The learning processes may be implemented by a fuzzy
neural network configured to form a look-up table for the fuzzy rule.
In yet another embodiment, the invention comprises a new physical measure
of control quality based on minimum production entropy and using this
measure for a fitness function of genetic algorithm in optimal control
system design. This method provides a local entropy feedback loop in the
control system. The entropy feedback loop provides for optimal control
structure design by relating stability of the plant (using a Lyapunov
function) and controllability of the plant (based on production entropy of
the control system). The control system is applicable to all control
systems, including, for example, control systems for mechanical systems,
bio-mechanical systems, robotics, electromechanical systems, etc.
BRIEF DESCRIPTION OF THE DRAWINGS
The advantages and features of the disclosed invention will readily be
appreciated by persons skilled in the art from the following detailed
description when read in conjunction with the drawings listed below.
FIG. 1 is a block diagram showing an example of an AI control method in the
prior art.
FIG. 2 is a block diagram showing an embodiment of an AI control method in
accordance with one aspect of the present invention.
FIG. 3A is a block diagram of an optimal control system.
FIG. 3B is a block diagram of a reduced control system.
FIG. 4A is an overview block diagram of a reduced control system that uses
entropy-based soft computing.
FIG. 4B is a detailed block diagram of a reduced control system that uses
entropy-based soft computing.
FIG. 5A is a system block diagram of a fuzzy logic controller.
FIG. 5B is a computing block diagram of the fuzzy logic controller shown in
FIG. 5A.
FIG. 6 is a diagram of an internal combustion piston engine with sensors.
FIG. 7 is a block diagram of a control system for controlling the internal
combustion piston engine shown in FIG. 6.
FIG. 8 is a schematic diagram of one half of an automobile suspension
system.
FIG. 9 is a block diagram of a control system for controlling the
automobile suspension system shown in FIG. 8.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT
A feedback control system is commonly used to control an output variable of
a process or plant in the face of some disturbance. Linear feedback
control systems typically use combinations of proportional feedback
control, integral feedback control, and derivative feedback control.
Feedback that is the sum of proportional plus integral plus derivative
feedback is often referred to as PID control. The Laplace transform of an
output u(s) of a PID controller is given by:
##EQU1##
In the above equation, G(s) is the transfer function of the PID controller,
e(s) is the controller input, u(s) is the controller output, k.sub.1 is
the coefficient for proportional feedback, k.sub.2 is; the coefficient for
integral feedback, and k.sub.3 is the coefficient for derivative feedback.
The coefficients k.sub.i may be represented by a coefficient vector K,
where K=[k.sub.1, k.sub.2, k.sub.3 ]. The vector K is commonly called a
Coefficient Gain Schedule (CGS). The values of the coefficients K used in
the linear PID control system are based on a dynamic model of the plant.
When the plant is unstable, nonlinear, and/or time-variant, then the
coefficients in K are often controlled by an AI control system.
FIG. 1 shows a typical prior art AI control system 100. An input y(t) of
the control system 100 is provided to a plus input of an adder 104 and an
output x(t) of a plant 110 is provided to a minus input of the adder 104.
An output of the adder 104 is provided as an error signal e(t) to an error
signal input of a PID controller 106. An output u(t) of the PID controller
106 is provided to a first input of an adder 108. A disturbance m(t) is
provided to a second input of the adder 108. An output u*(t) of the adder
108 is provided to an input of the plant 110. The plant 110 has a transfer
function H(s) and an output x(t), where x(t).revreaction.X(s) (where the
symbol.revreaction.denotes the Laplace transform) and X(s)=H(s)u*(s). An
output of the genetic algorithm 116 is provided to an input of a Fuzzy
logic Neural Network (FNN) 118 and an output of the fuzzy neural network
118 is provided to a Fuzzy Controller (FC) 120. An output of the fuzzy
controller 120 is a set of coefficients K, which are provided to a
coefficient input of the PID controller 106.
The error signal e(t) provided to the PID controller 106 is the difference
between the desired plant output value y(t) and the actual plant output
value x(t). The PID controller 106 is designed to minimize the error
represented by e(t) (the error being the difference between the desired
and actual output signal signals). The PID controller 106 minimizes the
error e(t) by generating an output signal u(t) which will move the output
signal x(t) from the plant 110 closer to the desired value. The genetic
algorithm 116, fuzzy neural network 118, and fuzzy controller 120 monitor
the error signal e(t) and modify the gain schedule K of the PID controller
106 in order to improve the operation of the PID controller 106.
The PID controller 106 constitutes a reverse model relative to the plant
110. The genetic algorithm 116 evolves an output signal a based on a
performance function .function.. Plural candidates for .alpha. are
produced and these candidates are paired according to which plural
chromosomes (parents) are produced. The chromosomes are evaluated and
sorted from best to worst by using the performance function .function..
After the evaluation for all parent chromosomes, good offspring
chromosomes are selected from among the plural parent chromosomes, and
some offspring chromosomes are randomly selected. The selected chromosomes
are crossed so as to produce the parent chromosomes for the next
generation. Mutation may also be provided. The second-generation parent
chromosomes are also evaluated (sorted) and go through the same
evolutionary process to produce the next-generation (i.e.,
third-generation) chromosomes. This evolutionary process is continued
until it reaches a predetermined generation or the evaluation function
.function. finds a chromosome with a certain value. The outputs of the
genetic algorithm are the chromosomes of the last generation. These
chromosomes become input information a provided to the fuzzy neural
network 118.
In the fuzzy neural network 118, a fuzzy rule to be used in the fuzzy
controller 120 is selected from a set of rules. The selected rule is
determined based on the input information a from the genetic algorithm
116. Using the selected rule, the fuzzy controller 120 generates a gain lo
schedule K for the PID controller 106. The vector coefficient gain
schedule K is provided to the PID controller 106 and thus adjusts the
operation of the PID controller 106 so that the PID controller 106 is
better able to minimize the error signal e(t).
Although the AI controller 100 is advantageous for accurate control in
regions near linearized equilibrium points, the accuracy deteriorates in
regions away from the linearized equilibrium points. Moreover, the AI
controller 100 is typically slow or even unable to catch up with changes
in the environment surrounding the plant 110. The PID controller 106 has a
linear transfer function G(s) and thus is based upon a linearized equation
of motion for the plant 110. Since the evaluation function .function. used
in the genetic algorithm 116 is only based on the information related to
the input e(t) of the linear PID controller 106, the controller 100 does
not solve the problem of poor controllability typically seen in
linearization models. Furthermore, the output results, both in the gain
schedule K and the output x(t) often fluctuate greatly, depending on the
nature of the performance functions used in the genetic algorithm 116. The
genetic algorithm 116 is a nonlinear optimizer that optimizes the
performance function .function.. As is the case with most optimizers, the
success or failure of the optimization often ultimately depends on the
selection of the performance function .function..
The present invention solves these and other problems by providing a new AI
control system 200 shown in FIG. 2. Unlike prior AI control systems, the
control system 200 is self-organizing and uses a new performance function
.function., which is based on the physical law of minimum entropy. An
input y(t) of the control system 200 is provided to a plus input of an
adder 204 and an output x(t) of a plant 210 is provided to a minus input
of the adder 204. An output of the adder 204 is provided as an error
signal e(t) to an error signal input of a PID controller 206 and to an
input of a fuzzy controller 220. An output u(t) of the PID controller 206
is provided to a first input of an adder 208 and to a first input of an
entropy calculator (EC) 214. A disturbance m(t) is provided to a second
input of the adder 208. An output u*(t) of the adder 208 is provided to am
input of the plant 210. The plant 210 has a transfer function H(s) and an
output x(t), such that X(s)=H(s)u*(s), where x(t).revreaction.X(s). The
output x(t) is provided to a second input of the entropy calculator 214
and to the minus input of the adder 204. An output of the entropy
calculator 214 is provided to an input of a genetic algorithm 216 and an
output of the genetic algorithm 216 is provided to an input of a Fuzzy
logic Neural Network (FNN) 218. An output of the fuzzy neural network 218
is provided to a rules selector input 222 of the fuzzy controller 220. A
Coefficient Gain Schedule (CGS) output 212 of the fuzzy controller 222 is
provided to a gain schedule input of the PID 206.
The combination of the genetic algorithm 216 and the entropy calculator 214
comprises a Simulation System of Control Quality 215. The combination of
the fuzzy neural network 218 and the fuzzy controller 220 comprises a
Fuzzy Logic Classifier System FLCS 219. The combination of the plant 210
and the adder 208 comprises a disturbed plant model 213. The disturbed
plant signal u*(t)=u(t)+m(t), and the disturbance m(t) are typically
unobservable.
The error signal e(t) provided to the PID controller 206 is the difference
between the desired plant output value y(t) and the actual plant output
value x(t). The PID controller 206 is designed to minimize the error
represented by e(t). The PID controller 206 minimizes the error e(t) by
generating an output signal u(t) which will move the output signal x(t)
from the plant 210 closer to the desired value. The fuzzy controller 220
monitors the error signal e(t) and modifies the gain schedule K of the PID
controller 206 according to a fuzzy control rule selected by the fuzzy
neural network 218.
The genetic algorithm 216 provides a teaching signal K.sub.T to the fuzzy
neural network 218. The teaching signal K.sub.T is a global optimum
solution of a coefficient gain schedule K generated by the genetic
algorithm 216. The PID controller 206 constitutes a reverse model relative
to the plant 210. The genetic algorithm 216 evolves an output signal a
based on a performance function .function.. Plural candidates for a are
produced and these candidates are paired by which plural chromosomes
(parents) are produced. The chromosomes are evaluated and sorted from best
to worst by using the performance function .function.. After the
evaluation for all parent chromosomes, good offspring chromosomes are
selected from among the plural parent chromosomes, and some offspring
chromosomes are randomly selected. The selected chromosomes are crossed so
as to produce the parent chromosomes for the next generation. Mutation is
also employed. The second-generation parent chromosomes are also evaluated
(sorted) and go through the same evolutionary process to produce the
next-generation (i.e., third-generation) chromosomes. This evolutionary
process is continued until it reaches a predetermined generation or the
evaluation function .function. finds a chromosome with a certain value.
Then, a component from a chromosome of the last generation becomes a last
output, i.e., input information .alpha. provided to the fuzzy neural
network 218.
In the fuzzy neural network 218, a filmy rule to be used in the fuzzy
controller 220 is selected from a set of rules. The selected rule is
determined based on the input information a from the genetic algorithm
216. Using the selected rule, the fuzzy controller 220 generates a gain
schedule K for the PID controller 206. This is provided to the PID
controller 206 and thus adjusts the operation of the PID controller 206 so
that the PID controller 206 is better able to minimize the error signal
e(t).
The fitness function .function. for the genetic algorithm is given by
##EQU2##
where
##EQU3##
The quantity dS.sub.u /dt represents the rate of entropy production in the
output x(t) of the plant 210. The quantity dS.sub.c /dt represents the
rate of entropy production in the output u(t) of the PID controller 206.
Entropy is a concept that originated in physics to characterize the heat,
or disorder, of a system. It can also be used to provide a measure of the
uncertainty of a collection of events, or, for a random variable, a
distribution of probabilities. The entropy function provides a measure of
the lack of information in the probability distribution. To illustrate,
assume that p(x) represents a probabilistic description of the known state
of a parameter, that p(x) is the probability that the parameter is equal
to z. If p(x) is uniform, then the parameter p is equally likely to hold
any value, and an observer will know little about the parameter p. In this
case, the entropy function is at its maximum. However, if one of the
elements of p(z) occurs with a probability of one, then an observer will
know the parameter p exactly and have complete information about p. In
this case, the entropy of p(x) is at its minimum possible value. Thus, by
providing a measure of uniformity, the entropy function allows
quantification of the information on a probability distribution.
It is possible to apply these entropy concepts to parameter recovery by
maximizing the entropy measure of a distribution of probabilities while
constraining the probabilities so that they satisfy a statistical model
given measured moments or data. Though this optimization, the distribution
that has the least possible information that is consistent with the data
may be found. In a sense, one is translating all of the information in the
data into the form of a probability distribution. Thus, the resultant
probability distribution contains only the information in the data without
imposing additional structure. In general, entropy techniques are used to
formulate the parameters to be recovered in terms of probability
distributions and to describe the data as constraints for the
optimization. Using entropy formulations, it is possible to perform a wide
range of estimations, address ill-posed problems, and combine information
from varied sources without having to impose strong distributional
assumptions.
Entropy-based control of the plant in FIG. 2 is based on obtaining the
difference between a time differentiation (dS.sub.u /dt) of the entropy of
the plant and a time differentiation (dS.sub.c /dt) of the entropy
provided to the plant from a low-level controller that controls the plant
210, and then evolving a control rule using a genetic algorithm. The time
derivative of the entropy is called the entropy production rate. The
genetic algorithm uses the difference between the entropy production rate
of the plant (dS.sub.u /dt and the entropy production rate of the
low-level controller (dS.sub.c /dt) as a performance function. Nonlinear
operation characteristics of the physical plant 210 are calculated by
using a Lyapunov function
The dynamic stability properties of the plant 210 near an equilibrium point
can be determined by use of Lyapunov functions. Let V(x) be a continuously
differentiable scalar function defined in a domain D.OR
right.R.sup..alpha. that contains the origin. The function V(x) is said to
be positive definite if V(0)=0 and V(x)>0 for x.noteq.0. The function V(x)
is said to be positive semidefinite if V(x).gtoreq.0 for all x. A function
V(x) is said to be negative definite or negative semidefinite if -V(x) is
positive definite or positive semidefinite, respectively. The derivative
of V along the trajectories x=.function.(x) is given by:
##EQU4##
where .differential.V/.differential.x is a row vector whose ith component
is .differential.V/.differential.x.sub.i and the components of the
n-dimensional vector .function.(x) are locally Lipschitz functions of x,
defined for all x in the domain D. The Lyapunov stability theorem states
that the origin is stable if there is a continuously differentiable
positive definite function V(x) so that V(x) is negative definite. A
function V(x) satisfying the conditions for stability is called a Lyapunov
function.
Calculation of the Lyapunov dynamic stability and entropy production for a
closed nonlinear mechanical system is demonstrated by using the
Holmes-Rand (Duffing-Van der Pol) nonlinear oscillator as an example. The
Holmes-Rand oscillator is described by the equation:
x+(.alpha.+.beta.x.sup.2)x-.gamma.x+x.sup.3 =0 (5)
where .alpha., 62 , and .gamma. are constant parameters. A Lyapunov
function for the Holmes-Rand oscillator is given by:
##EQU5##
where
##EQU6##
Entropy production d.sub.i S/dt for the Holmes-Rand oscillator is given by
the equation:
##EQU7##
Equation 5 can be rewritten as:
##EQU8##
After multiplying both sides of the above equation by x, then dV/dt can be
calculated as::
##EQU9##
where T is a normalizing factor.
An interrelation between a Lyapunov function and the entropy production in
an open dynamic system can be established by assuming a Lyapunov function
of the form
##EQU10##
where S=S.sub.u -S.sub.c, and the q.sub.i are generalized coordinates. It
is possible to introduce the entropy function S in the Lyapunov function V
because entropy S is also a scalar function of time. Differentiation of V
with respect to time gives:
##EQU11##
In this case, q.sub.i =.phi..sub.i (q.sub.i,.tau.,t), S=S.sub.u -S.sub.c,
S=S.sub.u -S.sub.c and thus:
##EQU12##
A special case occurs when .beta.=0 and the Holmes-Rand oscillator reduces
to a force-free Duffing oscillator, wherein:
##EQU13##
A Van der Pol oscillator is described by the equation:
x+(x.sup.2 -1)x+x=0 (14)
and the entropy production is given by:
##EQU14##
For a micro-mobile robot in fluid, a mechanical model is given by:
##EQU15##
where:
##EQU16##
Values for a particular micro-mobile robot are given in Table 1 below.
TABLE 1
Item Value Units
m.sub.1 1.6 .times. 10.sup.-7 kg
m.sub.2 1.4 .times. 10.sup.-6 kg
m.sub.3 2.4 .times. 10.sup.-6 kg
l.sub.1 2.0 .times. 10.sup.-3 m
l.sub.1 4.0 .times. 10.sup.-3 m
l.sub.3 4.0 .times. 10.sup.-3 m
K.sub.1 61.1 N/m
K.sub.2 13.7 N/m
K.sub.3 23.5 N/m
A.sub.1 4.0 .times. 10.sup.-6 m.sup.2
A.sub.2 2.4 .times. 10.sup.-6 m.sup.2
A.sub.3 4.0 .times. 10.sup.-6 m.sup.2
C.sub.d 1.12 --
.rho. 1000 Kg/m.sup.3
Entropy production for the micro-mobile robot is given by the equation:
##EQU17##
and the Lyapunov Function is given by:
##EQU18##
where S=S.sub.i -S.sub.c and S.sub.c is the entropy of a controller with
torque .tau..
The necessary and sufficient conditions for Lyapunov stability of a plant
is given by the relationship:
##EQU19##
According to the above equation, stability of a plant can be achieved with
"negentropy" S.sub.c (in the terminology used by Brillouin) where a change
of negentropy dS.sub.c /dt in the control system 206 is subtracted from a
change of entropy dS.sub.i /dt in the motion of the plant 210.
The robust AI control system 200 provides improved control of mechanical
systems in stochastic environments (e.g., active vibration control),
intelligent robotics and electro-mechanical systems (e.g., mobile robot
navigation, manipulators, collective mobile robot control), bio-mechanical
systems (e.g., power assist systems, control of artificial replaced organs
in medical systems as artificial lung ventilation), micro
electromechanical systems (e.g., micro robots in fluids), etc.
The genetic algorithm realizes the search of optimal controllers with a
simple structure using the principle of minimum entropy production. The
fuzzy neural network controller offers a more flexible structure of
controllers with a smaller torque, and the learning process produces less
entropy. The fuzzy neural network controller gives a more flexible
structure to controllers with smaller torque and the learning process
produces less entropy than a genetic analyzer alone. Thus, an instinct
mechanism produces less entropy than an intuition mechanism. However,
necessary time for achieving an optimal control with learning process on
fuzzy neural network (instinct) is larger than with the global search on
genetic algorithm (intuition).
Realization of coordinated action between the look-up tables of the fuzzy
controller 220 is accomplished by the genetic algorithm and the fuzzy
neural network. In particular, the structure 200 provides a multimode
fuzzy controller coupled with a linear or nonlinear neural network 218.
The control system 200 is a realization of a self-organizing AI control
system with intuition and instinct. In the adaptive controller 200, the
feedback gains of the PID controller 210 are charged according to the
quantum fuzzy logic, and approximate reasoning is provided by the use of
nonlinear dynamic motion equations.
The fuzzy tuning rules for the gains k.sub.i are shaped by the learning
system in the fuzzy neural network 218 with acceleration of fuzz rules on
the basis of global inputs provided by the genetic algorithm 216. The
control system 200 is thus a hierarchical, two-level control system that
is intelligent "in small." The lower (execution) level is provided by a
traditional PID controller 206, and the upper (coordination) level is
provided by a KB (with fuzzy inference module in the form of production
rules with different model of fuzzy implication) and fuzzification and
de-fuzzification components, respectively.
The genetic algorithm 216 simulates an intuition mechanism of choosing the
optimal structure of the PID controller 206 by using the fitness function,
which is the measure of the entropy production, and the evolution
function, which in this case is entropy.
Reduced Sensor Control Systems
In one embodiment, the entropy-based control system is applied to a reduced
control system wherein the number of sensors has been reduced from that of
an optimal system.
FIG. 3A shows an optimal control system x 302 controlling a plant 304. The
optimal control system x produces an output signal x. A group of m sensors
collect state information from the plant 304 and provide the state
information as feedback to the optimal control system x 302. The control
system x 302 is optimal not in the sense that it is perfect or best, but
rather, the control system x 302 is optimal in the sense that it provides
a desired control accuracy for controlling the output of the plant 304.
FIG. 3B shows a reduced control system y 312 controlling the plant 304. The
reduced control system y 312 provides an output signal y. A group of n
sensors (where n<m) collect state information from the plant 304 and
provide the state information as feedback to the reduced control system x
312. The reduced control system 312 is reduced because it receives
information from fewer sensors than the optimal control system 302 (n is
less than m).
The reduced control system 312 is configured to use smart simulation
techniques for controlling the plant 304 despite the reduction in the
number of sensor number without significant loss of control quality
(accuracy) as compared to the optimal control system 302. The plant 304
may be either stable or unstable, and the control system 312 can use
either a complete model of the plant 304 or a partial model of the plant
304.
The reduced control system 312 provides reliability, sufficient accuracy,
robustness of control, stability of itself and the plant 304, and lower
cost (due, in part, to the reduced number of sensors needed for control).
The reduced control system 312 is useful for robust intelligent control
systems for plants such as, for example, suspension system, engines,
helicopters, ships, robotics, micro electromechanical systems, etc.
In some embodiments, reduction in the number of sensors may be used, for
example, to integrate sensors as intelligent systems (e.g.,
"sensor-actuator-microprocessor" with the preliminary data processing),
and for simulation of robust intelligent control system with a reduced
number of sensors.
The amount of information in the reduced output signal y (with n sensors)
is preferably close to the amount of information in the optimal output
signal. In other words, the reduced (variable) output y is similar to the
optimal output x, such that (x.apprxeq.y). The estimated accuracy of the
output y as related to x is M[(x-y).sup.2 ].ltoreq..epsilon..sup.2.
In some embodiments, the reduced control system 312 is adapted to minimize
the entropy production in the control system 312 and in the plant 304.
The general structure of a reduced control system is shown in FIGS. 4A and
4B. FIG. 4A is a block diagram showing a reduced control system 480 and an
optimal control system 420. The optimal control system 420, together with
an optimizer 440 and a sensor compensator 460 are used to teach the
reduced control system 480. In FIG. 4A, a desired signal (representing the
desired output) is provided to an input of the optimal control system 420
and to an input of the reduced control system 480. The optimal control
system 420, having m sensors, provides an output sensor signal x.sub.b and
an optimal control signal x.sub.a. A reduced control system 480 provides
an output sensor signal y.sub.b and a reduced control signal y.sub.a. The
signals x.sub.b and y.sub.b includes data from k sensors where
k.ltoreq.m-n. The k sensors typically being the sensors that are not
common between the sensor systems 422 and 482. The signals x.sub.b and
y.sub.b are provided to first and second inputs of a subtractor 491. An
output of the subtractor 491 is a signal .epsilon..sub.b, where
.epsilon..sub.b =x.sub.b -y.sub.b. The signal .epsilon..sub.b is provided
to a sensor input of a sensor compensator 460. The signals x.sub.a and
y.sub.a are provided to first and second inputs of a subtractor 490. An
output of the subtractor 490 is a signal .epsilon..sub.a where
.epsilon..sub.a =x.sub.a -y.sub.a. The signal .epsilon..sub.a is provided
to a control signal input of the sensor compensator 460. A control
information output of the sensor compensator 460 is provided to a control
information input of the optimizer 440. A sensor information output of the
sensor compensator 460 is provided to a sensor information input of the
optimizer 440. A sensor signal 483 from the reduced control system 480 is
also provided to an input of the optimizer 440. An output of the optimizer
440 provides a teaching signal 443 to an input of the reduced control
system 480.
In the description that follows, off-line mode typically refers to a
calibration mode, wherein the control object 428 (and the control object
488) is run with an optimal set of set of m sensors. In one embodiment,
the off-line mode is rum at the factory or repair center where the
additional sensors (i.e., the sensors that are in the m set but not in the
n set) are used to train the FNN1426 and the FNN2486. The online mode
typically refers to an operational mode (e.g., normal mode) where the
system is run with only the n set of sensors.
FIG. 4B is a block diagram showing details of the blocks in FIG. 4A. In
FIG. 4B, the output signal x.sub.k is provided by an output of a sensor
set m 422 having m sensors given by m=k+k.sub.1. The information from the
sensor system m 422 is a signal (or group of signals) having optimal
information content I.sub.1. In other words, the information I.sub.1 is
the information from the complete set of m sensors in the sensor system
422. The output signal x, is provided by an output of a control unit 425.
The control signal x.sub.a is provided to an input of a control object
428. An output of the control object 428 is provided to an input of the
sensor system 422. Information I.sub.k from the set k of sensors is
provided to an online learning input of a fuzzy neural network (FNN1) 426
and to an input of a first Genetic Algorithm (GA1) 427. Information
I.sub.k1 from the set of sensors k.sub.1 in the sensor system 422 is
provided to an input of a control object model 424. An off-line tuning
signal output from the algorithm GA1427 is provided to an off-line tuning
signal input of the FNN1426. A control output from the FNN 426 is the
control signal x.sub.a, which is provided to a control input of the
control object 428. The control object model 424 and the FNN 426 together
comprise an optimal fuzzy control unit 425.
Also in FIG. 4B, the sensor compensator 460 includes a multiplier 462, a
multiplier 466, an information calculator 468, and an information
calculator 464. The multiplier 462 an(i the information calculator 464 are
used in online (normal) mode. The multiplier 466 and the information
calculator 468 are provided for off-line checking.
The signal .epsilon..sub.a from the output of the adder 490 is provided to
a first input of the multiplier 462 and to a second input of the
multiplier 462. An output of the multiplier 462, being a signal
.epsilon..sub.a.sup.2, is provided to an input of the information
calculator 464. The information calculator 41A computes H.sub.a
(y).ltoreq.I(x.sub.a,y.sub.a). An output of the information calculator 464
is an information criteria for accuracy and reliability,
I(x.sub.a,y.sub.a).fwdarw.max.
The signal .epsilon..sub.b from the output of the adder 491 is provided to
a first input of the multiplier 466 and to a second input of the
multiplier 466. An output of the multiplier 466, being a signal
.epsilon..sub.b.sup.2, is provided to an input of the information
calculator 468. The information calculator 468 computes H.sub.b
(y).ltoreq.I(x.sub.b,y.sub.b). An output of the information calculator 468
is an information criteria for accuracy and reliability, I(x.sub.b
y.sub.b).fwdarw.max.
The optimizer 440 includes a second Genetic Algorithm (GA2) 444 and an
entropy model 442. The signal I(x.sub.a,y.sub.1).fwdarw.max from the
information calculator 464 is provided to a first input of the algorithm
(GA2) 444 in the optimizer 440. An entropy signal S.fwdarw.min is provided
from an output of the thermodynamic model 442 to a second input of the
algorithm GA2444. The signal I(x.sub.b,y.sub.b).fwdarw.max from the
information calculator 468 is provided to a third input of the algorithm
(GA2) 444 in the optimizer 440.
The signals I(x.sub.a,y.sub.a).fwdarw.max and I(x.sub.b,y.sub.b).fwdarw.max
provided to the first and third inputs of the algorithm (GA2) 444 are
information criteria, and the entropy signal S(k.sub.2).fwdarw.min
provided to the second input of the algorithm GA2444 is a physical
criteria based on entropy. An output of the algorithm GA2444 is a teaching
signal for a FNN2486 described below.
The reduced control system 480 includes a reduced sensor set 482, a control
object model 484, the FNN2486, and a control object 488. When run in a
special off-line checking (verification) mode, the sensor system 482 also
includes the set of sensors k. The control object model 484 and the
FNN2486 together comprise a reduced fuzzy control unit 485. An output of
the control object 488 is provided to an input of the sensor set 482. An
I.sub.2 output of the sensor set 482 contains information from the set n
of sensors, where n=(k.sub.1 +k.sub.2)<m, such that I.sub.2 <I.sub.1. The
information I.sub.2 is provided to a tuning input of the FNN2486, to an
input of the control object model 484, and to an input of the entropy
model 442. The teaching signal 443 from the algorithm GA2444 is provided
to a teaching signal input of the FNN2486. A control output from the
FNN2486 is the signal y.sub.a, which is provided to a control input of the
control object 488.
The control object models 424 and 484 may be either full or partial models.
Although FIGS. 4A and 4B show the optimal system 420 and the reduced
system 480 as separate systems, typically the systems 420 and 480 are the
same system. The system 480 is "created" from the system 420 by removing
the extra sensors and training the neural network. Thus, typically, the
control object models 424 and 484 are the same. The control object 428 and
488 are also typically the same.
FIG. 4B shows arrow an off-line tuning arrow mark from GA1427 to the
FNN1426 and from the GA2444 to the FNN2486. FIG. 4B also shows an online
learning arrow mark 429 from the sensor system 422 to the FNN1426. The
Tuning of the GA2444 means to change a set of bonding coefficients in the
FNN2486. The bonding coefficients are changed (using, for example, an
iterative or trial and error process) so that I(x,y) trends toward a
maximum and S trends toward a minimum. In other words, the information of
the coded set of the coefficients is sent to the FNN2486 from the GA2444
as I(x,y) and S are evaluated. Typically, the bonding coefficients in the
FNN2486 are tuned in off-line mode at a factory or service center.
The teaching signal 429 is a signal that operates on the FNN1426 during
operation of the optimal control system 420 with an optimal control set.
Typically, the teaching signal 429) is provided by sensors that are not
used with the reduced control system 480 when the reduced control system
is operating in online mode. The GA1427 tunes the FNN1426 during off-line
mode. The signal lines associated with x.sub.b and y.sub.b are dashed to
indicate that the x.sub.b and y.sub.b signals are typically used only
during a special off-line checking (i.e., verification) mode. During the
verification mode, the reduced control system is run with an optimal set
of sensors. The additional sensor information is provided to the optimizer
440 and the optimizer 440 verifies that the reduced control system 480
operates with the desired, almost optimal, accuracy.
For stable and unstable control objects with a non-linear dissipative
mathematical model description and reduced number of sensors (or a
different set of sensors) the control system design is connected with
calculation of an output accuracy of the control object and reliability of
the control system according to the information criteria
I(x.sub.a,y.sub.a).fwdarw.max and I(x.sub.b,y.sub.b).fwdarw.max. Control
system design is also connected with the checking of stability and
robustness of the control system and the control object according to the
physical criteria S(k.sub.2).fwdarw.min.
In a first step, the genetic algorithm GA2444 with a fitness function
described as the maximum of mutual information between and optimal control
signals x.sub.a and a reduced control signal y.sub.a is used to develop
the teaching signal 443 for the fuzzy neural network FNN2486 in an
off-line simulation. The fuzzy neural network FNN2486 is realized using
the learning process with back-propagation for adaptation to the teaching
signal, and to develop a lookup-table :for changing the parameters of a
PID-controller in the controller 485. This provides sufficient conditions
for achieving the desired control reliability with sufficient accuracy.
In a second step, the genetic algorithm GA2444, with a fitness function
described as the minimum entropy, S, production rate dS/dt, (calculated
according to a mathematical model of the control object 488), is used to
realize a node correction lookup-table in the FNN2486. This approach
provides stability and robustness of the reduced control system 480 with
reliable and sufficient accuracy of control. This provides a sufficient
condition of design for the robust intelligent control system with the
reduced number of sensors.
The first and second steps above need not be done in the listed order or
sequentially. In the simulation of unstable objects, both steps are
preferably done in parallel using the fitness function as the sum of the
physical and the information criteria.
After the simulation of a lookup-table for the FNN2486, the mathematical
model of control object 484 is changed on the sensor system 482 for
checking the qualitative characteristics between the reduced control
system with the reduced number of sensors and the reduced control system
with a full complement (optimum number) of sensors. Parallel optimization
in the (GA2444 with two fitness functions is used to realize the global
correction of a lookup-table in, the FNN2486.
The entropy model 442 extracts data from the sensor information I.sub.2 to
help determine-, the desired number of sensors for measurement and control
of the control object 488.
FIGS. 4A and 4B show the general case when the reduction of sensors
excludes the measurement sensors in the output of the control object and
the calculation comparison of control signal on information criteria is
possible. The sensor compensator 460 computes the information criteria as
a maximum of mutual information between the two control signals x.sub.a
and y.sub.a (used as the first fitness function for the GA2444). The
entropy model 442 present the physical criteria as a minimum production
entropy (used as the second fitness function in the GA2444) using tie
information from sensors 482. The output of the GA2444 is the teaching
signal 447 for the FNN2486 that is used on-line to develop the reduced
control signal y.sub.a such that the quality of the reduced control signal
y.sub.b is similar to the quality of the optimal control signal x.sub.a.
Thus, the optimizer 440 provides stability and robustness of control
(using the physical criteria), and reliability with sufficient accuracy
(using the information criteria).
With off-line checking, the optimizer 440 provides correction of the
control signal y.sub.a from the FNN2486 using new information criteria
Since the information is additive, it is possible to do the
online/off-line steps in sequence, or in parallel. In off-line checking,
the sensor system 482 typically uses all of the sensors only for checking
of the quality and correction of the control signal y.sub.a. This approach
provides the desired reliability and quality of control, even when the
control object 488 is unstable.
For the reduced sensor system 482 (with n sensors) the FNN2486 preferably
uses a learning and adaptation process instead of a Fuzzy Controller (FC)
algorithm. FIG. 5 illustrates the similarity between a FC 501 and a FNN
540.
As shown in FIG. 5A, the structure of the FNN 540 is similar to the
structure of the FC 501. FIG. 5A shows a PID controller 503 having a first
input 535 to receive a desired sensor output value and a second input 536
to receive an actual sensor output value 536. The desired sensor output
value is provided to a plus input of an adder 520 and the actual sensor
output value 536 is provided to a minus input of the adder 520. An output
of the adder 520 is provided to a proportional output 531, to an input of
an integrator 521, and to an input of a differentiator 522. An output of
the integrator is provided to an integral output 532, and an output of the
differentiator is provided to a differentiated output 533.
The Fuzzy logic Controller (FC) 501 includes a fuzzification interface 504
(shown in FIG. 5B), a Knowledge Base (KB) 502, a fuzzy logic rules
calculator 506, and a defuzzification interface 508. The PID outputs
531-533 are provided to PID inputs of the fuzzification interface 504. An
output of the fuzzification interface 504 is provided to a first input of
the fuzzy logic rules calculator 506. An output of the KB 502 is provided
to a second input of the fuzzy logic rules calculator 506. An output of
the fuzzy logic rules calculator 506 is provided to an input of the
defuzzification interface 508, and an output of the defuzzification
interface 508 is provided as a control output U.sub.FC of the FC 501.
The Fuzzy Neural Network includes six neuron layers 550-555. The neuron
layer 550 is an input layer and has three neurons, corresponding to the
three PID output signals 531-533. The neuron layer 550 is an output layer
and includes a single neuron. The neuron layers 551-553 are hidden layers
(e.g., not visible from either the input or the output). The neuron layer
551 has six neurons in two groups of three, where each group is associated
with one of the three input neurons in the neuron layer 550. The neuron
layer 552 includes eight neurons. Each neuron, in the neuron layer 552 has
three inputs, where each input is associated with one of the three groups
from the neuron layer 551. The neuron layer 553 has eight neurons. Each
neuron in the neuron layer 553 receives input from all of the neurons in
the neuron layer 552. The neuron layer 554 has is a visible layer having 8
neurons. Each neuron in the neuron layer 554 has four inputs and receives
data from one of the neurons in the neuron layer 553 (i.e., the first
neuron in the neuron layer 554 receives input from the first neuron in the
neuron layer 553, the second from the second, etc.). Each neuron in the
neuron layer 554 also receives input from all of the PID input signals
531-533. The output neuron layer 555 is a single neuron with eight inputs
and the single neuron receives data from all of the neurons in the layer
554 (a generalized Takagi-Sugeno FC models.
The number of hidden layers in the FNN 540 typically corresponds to the
number of elements in the FC 501. One distinction between the FC 501 and
the FNN 540 lies in the way in which the knowledge base 502 is formed. In
the FC 501, the parameters and membership functions are typically fixed
and do not change during of control. By contrast, the FNN 540 uses the
learning and adaptation processes, such that when running on-line (e.g.,
normal model) the FNN 540 changes the parameters of the membership
function in the "IF . . . THEN" portions of the production rules.
The entropy-based control system in FIGS. 4A and 4B achieves a desired
control accuracy the output signal of a reduced control system y, as
related to an optimal system x, by using the statistical criteria of the
accuracy given above as:
M[(x-y).sup.2 ].ltoreq..epsilon..sup.2 (23)
The optimal output signal x has a constant dispersion .sigma..sub.x.sup.2
and the variable output signal y as a variable dispersion
.sigma..sub.y.sup.2. For a desired control accuracy .epsilon., the
.epsilon.-entropy H.sub..epsilon. (y) is a measure of the uncertainty in
the measurement of the process y as related to the process x, where the
processes x and y differ from each other in some metric by .epsilon..
The .epsilon.-entropy H.sub..epsilon. (y) can be calculated as
##EQU20##
where H(y) is the entropy of the process y and n is a number of point
measurements in the process y. Further,
##EQU21##
where p.sub.k is the probability that one of the output values of the
process y is equal to the value of the output values of the process x.
In general, the form of the .epsilon.-entropy is defined as
##EQU22##
with the accuracy M[(x-y).sup.2 ].ltoreq..epsilon..sup.2.
In general, the amount of information, I(x,y), is given by
##EQU23##
When x and y are Gaussian process, then
##EQU24##
where
##EQU25##
The amount of information, given by I(x,y), satisfies the inequality
I(x,y).gtoreq.H.sub..epsilon. (y) (29)
and, it may be seen that
M[(x-y).sup.2 ].gtoreq.[.epsilon..sub.y (I(x,y))].sup.2 (30)
where the function .epsilon..sub.y (I(x,y)) is the inverse of
H.sub..epsilon. (y)
H.sub..epsilon..sub..sub.y .sub.(H) (y)=H (31)
and thus
.epsilon..sub.y (I(x,y)).ltoreq..epsilon..sub.y (H.sub..epsilon.
(y)=M[(x-y).sup.2 ] (32)
Thus, it is possible to bound the mean-square error M[(x-y).sup.2
].ltoreq..epsilon..sub.2 from below by the known amount of information
I(x,y).
The calculation of
##EQU26##
determines the reliability of the measurements of the process y, as related
to the optimal process x, with the accuracy
.epsilon..sub.y (I(x,y)).ltoreq.M[(x-y).sup.2 ].ltoreq..epsilon..sup.2
(34)
In equation (28), the amount of information I(x,y) in the process y as
related to the process will have a maximal value I.sub.max when
.sigma..sub.y.sup.2.fwdarw..sigma..sub.y.sup.2. Moreover, at I.sub.max,
the probability p(x,y) of the similarity between the processes x and y is
(p(y.apprxeq.x).fwdarw.1), and thus, the minimum value of .epsilon..sup.2
is observed. The calculation of equation (33) for the control process y
compensates for the loss of information, and minimized the statistical
risk, caused by the reduction in the number of sensors.
Engine Control
In one embodiment, the reduced sensor, entropy-based, control system is
applied to an engine control system, such as, for example, an internal
combustion engine, a piston engine, a diesel engine, a jet engine, a
gas-turbine engine, a rocket engine, etc. FIG. 6 shows an internal
combustion piston engine having four sensors, an intake air temperature
sensor 602, a water temperature sensor 604, a crank angle sensor 606, and
an air-fuel ratio sensor 608. The air temperature sensor 602 measures the
air temperature in an intake manifold 620. A fuel injector 628 provides
fuel to the air in the intake manifold 620. The intake manifold 620
provides air and fuel to a combustion chamber 622. Burning of the fuel-air
mixture in the combustion chamber 622 drives a piston 628. The piston 628
is connected to a crank 626 such that movement of the piston 628 turns the
crank 626. The crank angle sensor 606 measures the rotational position of
the crank 626. The water temperature sensor measures the temperature of
water in a water jacket 630 surrounding the combustion chamber 622 and the
piston 628. Exhaust gasses from the combustion chamber 622 are provided to
an exhaust manifold 624 and the air-fuel ratio sensor 608 measures the
ratio of air to fuel in the exhaust gases.
FIG. 7 is a block diagram showing a reduced control system 780 and an
optimal control system 720. The optimal control system 720, together with
an optimizer 740 and a sensor compensator 760 are used to teach the
reduced control system 780. In FIG. 7, a desired signal (representing the
desired engine output) is provided to an input of the optimal control
system 720 and to an input of the reduced control system 780. The optimal
control system 720, having five sensors, provides an optimal control
signal x.sub.a and an sensor output signal x.sub.b. A reduced control
system 780 provides a reduced control output signal y.sub.a, and an output
sensor signal y.sub.b. The signals x.sub.b and y.sub.b include data from
the A/F sensor 608. The signals x.sub.b and y.sub.b are provided to first
and second inputs of a subtractor 791. An output of the subtractor 791 is
a signal .epsilon..sub.b where .epsilon..sub.b =x.sub.b -y.sub.b. The
signal .epsilon..sub.b is provided to a sensor input of a sensor
compensator 760. The signals x.sub.a and y.sub.a are provided to first and
second inputs of a subtractor 790. An output of the subtractor 790 is a
signal .epsilon..sub.a where .epsilon.=x.sub.a -y.sub.a. The signal
.epsilon..sub.a is provided to a control signal input of the sensor
compensator 760. A control information output of the sensor compensator
760 is provided to a control information input of the optimizer 740. A
sensor information output of the sensor compensator 760 is provided to a
sensor information input of the optimizer 740. A sensor signal 783 from
the reduced control system 780 is also provided to an input of the
optimizer 740. An output of the optimizer 740 provides a teaching signal
747 to an input of the reduced control system 780.
The output signal x.sub.b is provided by an output of a sensor system 722
having five sensors, including, the intake air temperature sensor 602, the
water temperature sensor 604, the crank angle sensor 606, the pressure
sensor 607, and the air-fuel ratio (A/F) sensor 608. The information from
the sensor system 722 is a group of signals having optimal information
content I.sub.1. In other words, the information I.sub.1 is the
information from the complete set of five sensors in the sensor system
722.
The output signal x.sub.a is provided by an output of a control unit 725.
The control signal x.sub.a is provided to an input of an engine 728. An
output of the engine 728 is provided to an input of the sensor system 722.
Information I.sub.k from the A/F sensor 608 is provided to an online
learning input of a fuzzy neural network (FNN) 726 and to an input of a
first Genetic Algorithm (GA1) 727. Information I.sub.k1 from the set of
four sensors excluding the A/F sensor 608 is provided to an input of an
engine model 724. An off-line tuning signal output from the algorithm
GA1727 is provided to an off-line tuning signal input of the FNN 726. A
control output from the FNN 726 is a fuel injector the control signal
U.sub.1, which is provided to a control input of the engine 728. The
signal U.sub.1 is also the signal x.sub.a. The engine model 724 and the
FNN 726 together comprise an optimal control unit 725.
The sensor compensator 760 includes a multiplier 762, a multiplier 766, an
information calculator 768, and an information calculator 764. The
multiplier 762 and the information calculator 764 are used in online
(normal) mode. The multiplier 766 and the information calculator 768 are
provided for off-line checking.
The signal .epsilon..sub.a from the output of the adder 790 is provided to
a first input of the multiplier 762 and to a second input of the
multiplier 762. An output of the multiplier 762, being a signal
.epsilon..sub.a.sup.2, is provided to an input of the information
calculator 764. The information calculator 764 computes H.sub.a
(y).ltoreq.I(x.sub.a,y.sub.a). An output of the information calculator 764
is an information criteria for accuracy and reliability,
I(x.sub.a,y.sub.a).fwdarw.max.
The signal .epsilon..sub.b from the output of the adder 791 is provided to
a first input of the multiplier 766 and to a second input of the
multiplier 766. An output of the multiplier 764, being a signal
.epsilon..sub.b.sup.2, is provided to an input of the information
calculator 768. The information calculator 768 computes H.sub.b
(y).ltoreq.I(x.sub.b,y.sub.b). An output of the information calculator 768
is an information criteria for accuracy and reliability,
I(x.sub.b,y.sub.b).fwdarw.max.
The optimizer 740 includes a second Genetic Algorithm (GA2) 744 and a
thermodynamic (entropy) model 742. The signal
I(x.sub.a,y.sub.b).fwdarw.max from the information calculator 764 is
provided to a first input of the algorithm (GA2) 744 in the optimizer 740.
An entropy signal S.fwdarw.min is provided from an output of the
thermodynamic model 742 to a second input of the algorithm GA2744. The
signal I(x.sub.b,y.sub.b).fwdarw.max from the information calculator 768
is provided to a third input of the algorithm (GA2) 744 in the optimizer
740.
The signals I(x.sub.a,y.sub.a).fwdarw.max and I(x.sub.b,y.sub.b).fwdarw.max
provided to the first and third inputs of the algorithm (GA2) 744 are
information criteria, and the entropy signal S(k.sub.2).fwdarw.min
provided to the second input of the algorithm GA2744 is a physical
criteria based on entropy. An output of the algorithm GA2744 is a teaching
signal for the FNN 786.
The reduced control system 780 includes a reduced sensor system 782, an
engine model 744, the FNN 786, and an engine 788. The reduced sensor
system 782 includes all of the engine sensors in the sensor system 722
except the A/F sensor 608. When run in a special off-line checking mode,
the sensor system 782 also includes the A/F sensor 608. The engine model
744 and the FNN 786 together comprise a reduced control unit 785. An
output of the engine 788 is provided to an input of the sensor set 782. An
I.sub.2 output of the sensor set 782 contains information from four engine
sensors, such that I.sub.2 <I.sub.1. The information I.sub.2 is provided
to an input of the control object model 784, and to an input of the
thermodynamic model 742. The teaching signal 747 from the algorithm GA2744
is provided to a teaching signal input of the FNN 786. A control output
from the FNN 786 is an injector control signal U.sub.2, which is also the
signal y.sub.a.
Operation of the system shown in FIG. 7 is in many respects similar to the
operation of the system shown in FIGS. 4A, 4B, and 5B.
The thermodynamic model 742 is built using the thermodynamic relationship
between entropy production and temperature information from the water
temperature sensor 604 (T.sup.W) and the air temperature sensor 602
(T.sup.A). The entropy production S(T.sup.W,T.sup.A), is calculated using
the relationship
##EQU27##
where .DELTA..tau. is the duration of a finite process.
The engine model 744 is analyzed from a thermodynamic point of view as a
steady-state model where the maximum work, which is equal to the minimum
entropy production, is delivered from a finite resource fluid and a bath.
Moreover, the engine model is analyzed as a dissipative, finite-time,
generalization of the evolutionary Carnot problem in which the temperature
driving force between two interacting subsystems varies with the constant
time .DELTA..tau. (a reversible cycle).
The presence of reversible cycles fixes, automatically, the first-law
efficiency of each infinitesimal stage at the Carnot level
##EQU28##
where T is the instantaneous temperature of a finite resource (such as the
water temperature) and T.sup.e is the temperature of the environment or an
infinite bath (such as the air temperature). The unit mass of a resource
releases the heat dq=-cdT, where c is the specific heat. The classical
rate and duration-dependent function of available energy (dissipative
energy) E.sub.x follows by integration of the product
##EQU29##
between the limits T and T.sup.e. The integration yields the well known
expression
##EQU30##
The problem becomes non-trivial in the case where a finite-rate process is
considered, because in the finite-rate process, the efficiency differs
from the Carnot efficiency. The finite-rate efficiency should be
determined before the integration of the product -cdT can be evaluated.
The integration then leads to a generalized available energy associated
with the extremal release of the mechanical work in a finite time.
The local efficiency of an infinitesimal unit of the finite-rate process is
##EQU31##
where Q.sub.1 is the cumulative heat flux flowing from the upper reservoir.
A flux Q.sub.2 is the cumulative heat flux flowing to the lower reservoir.
While this local first-law efficiency is still described by the Carnot
formula
##EQU32##
were T.sub.1.ltoreq.T'.sub.1.ltoreq.T'.sub.2.ltoreq.T.sub.2, this
efficiency is nonetheless lower than the efficiency of the unit working
between the boundary temperatures T.sup.1 and T.sub.2=T.sup.e , as the
former applies to the intermediate temperatures T.sub.1 ' and T.sub.2 '.
By solving equation (40) along with the reversible entropy balance of the
Carnot differential subsystem
##EQU33##
one obtains the primed temperatures as functions of the variables T.sub.1,
T.sub.2 =T.sup.e and .eta.. In equation (41), the parameters .gamma..sub.1
, and .gamma..sub.2 are the partial conductances, and thus link the heat
sources with the working fluid of the engine at high and low temperatures.
The differentials of the parameters .gamma. and .gamma..sub.2 can be
expressed as d.gamma..sub.1 =.alpha..sub.i dA.sub.i, for i=1,2, where the
values .alpha..sub.i are the heat transfer coefficients and the values
dA.sub.i are the upper and lower exchange surface areas.
The associated driving heat flux dQ.sub.1 =dy.sub.1 (T.sub.1 -T'.sub.1)) is
then found in the form
##EQU34##
from which the efficiency-power characteristic follows in the form
##EQU35##
From the energy balance of the driving fluid, the heat power variable
Q.sub.1 satisfies dQ=-GcdT, where dT is the differential temperature drop
of the driving fluid, G is the finite-mass flux of the driving fluid whose
constant specific heat capacity equals c. Using the above definitions, and
the differential heat balance of the driving fluid, the control term
dQ.sub.1 /d.gamma. from equation (43) may be written (with the subscript 1
omitted) as
##EQU36##
The negative of the derivative dQ/d.gamma. is the control variable u of the
process. In short, the above equation says that u=T, or that the control
variable u equals the rate of the temperature change with respect to the
non-dimensional time .tau.. Thus, equation (43) becomes the simple,
finite-rate generalization of the Carnot formula
##EQU37##
When T>T.sup.e then the derivative T is negative for the engine mode. This
is because the driving fluid releases the energy to the engine as part of
the work production.
Work Functionals for an Infinite Sequence of Infinitesimal Processes
The cumulative power delivered per unit fluid flow WIG is obtained by the
integration of the product of n and dQ/G=-cdT between an arbitrary initial
temperature T.sup.i and a final temperature T.sup.f of the fluid. This
integration yields the specific work of the flowing fluid in the form of
the functional
##EQU38##
The notation [T.sup.i,T.sup.f ] means the passage of the vector
T.tbd.(T,.tau.) from its initial state to its final state. For the above
functional, the work maximization problem can be stated for the engine
mode of the process by
##EQU39##
In equation (47), the function L(T,T) is the Lagrangian. The above
Lagrangian functional represents the total power per unit mass flux of the
fluid, which is the quantity of the specific work dimension, and thus the
direct relation to the specific energy of the fluid flow. In the
quasi-static limit of vanishing rates, where dT/d.tau.=0, the above work
functional represents the change of the classical energy according to
equation (38).
For the engine mode of the process, the dissipative energy itself is
obtained as the maximum of the functional in equation (47), with the
integration limits T=T.sup.e and T.sup.f =T.
An alternative form of the specific work can be written as the functional
##EQU40##
in which the first term is the classical "reversible" term and the second
term is the product of the equilibrium temperature and the entropy
production S, where
##EQU41##
The entropy generation rate is referred here to the unit mass flux of the
driving fluid, and thus to the specific entropy dimension of the quantity
S. The quantity S is distinguished from the specific entropy of the
driving fluid, s in equation (38). The entropy generation S is the
quadratic function of the process rate, u=dT/d.tau., only in the case when
the work is not produced, corresponding to the vanishing efficiency
.eta.=0 or the equality T+T=T.sup.e. For the activity heat transfer
(non-vanishing .eta.) case, the entropy production appears to be a
non-quadratic function of rates, represented by the integrand of equation
(48). Applying the maximum operation for the functional in equation (48)
at the fixed temperatures and times, it is seen that the role of the first
term (i.e., the potential term) is inessential, and the problem of the
maximum released work is equivalent to the associated problem of the
minimum entropy production. This confirms the role of the entropy
generation minimization in the context of the extremum work problem. The
consequence of the conclusion is that a problem of the extremal work and
an associated fixed-end problem of the minimum entropy generation have the
same solutions.
Hamilton-Jacobi Approach to Minimum Entropy Generation
The work extremization problems can be broken down to the variational
calculus for the Lagrangian
##EQU42##
The Euler-Lagrangian equations for the problem of extremal work and the
minimum entropy production lead to a second-order differential equation
TT-T.sup.c =0 (51)
which characterizes the optimal trajectories of all considered processes. A
first integral of the above equation provides a formal analog of the
mechanical energy as
##EQU43##
An equation for the optimal temperature flow from the condition E=h in
equation (38) is given by
##EQU44##
The power of dynamic programming (DP) methods when applied to problems of
this sort lies in the fact that regardless of local constraints, functions
satisfy an equation of the Hamilton-Jacobi-Bellman (HJB-equation) with the
same state variables as those for the unconstrained problem. Only the
numerical values of the optimizing control sets and the optimal
performance functions differ in constrained and unconstrained cases.
The engine problem may be correctly described by a backward HJB-equation.
The backward HJB-equation is associated with the optimal work or energy as
an optimal integral function I defined on the initial states (i.e.,
temperatures), and, accordingly, refers to the engine mode or process
approaching an equilibrium. The function I is often called the optimal
performance function for the work integral.
The maximum work delivery (whether constrained or unconstrained) is
governed by the characteristic function
##EQU45##
The quantity I in equation (54) describes the extremal value of the work
integral W(T.sup.i, T.sup.f) It characterizes the extremal value of the
work released for the prescribed temperatures T.sup.i and T.sup.f when the
total process duration is .tau..sup..eta. -.tau..sup.i. While the
knowledge of the characteristic function I only is sufficient for a
description of the extremal properties of the problem, other functions of
this sort are nonetheless useful for characterization of the problem.
Here the problem is transformed into the equivalent problem in which one
seeks the maximum of the final work coordinate x.sup.f.sub.o
=W.sup..function. for the system described by the following set of
differential equations
##EQU46##
The state of the system described by equations (55) is described by the
enlarged state vector X=(x.sub.0, x.sub.1, x.sub.2) which is composed of
the three state coordinates x.sub.0 =W, x.sub.1 =T, and x.sub.2 =.tau.. At
this point, it is convenient to introduce a set of optimal performance
functions .THETA..sup.i and .THETA..sup.f corresponding to the initial and
final work coordinates, respectively. The function .THETA..sup.i works in
a space of one dimension larger than I, involves the work coordinate
x.sub.0 =W, and is defined as
max W.sup..function..tbd..THETA..sup.i
(W.sup.i,.tau..sup.i,T.sup.i,.tau..sup..function.,T.sup..function.)=W.sup.
i +I(.tau..sup.i,T.sup.i,.tau..sup..function.,T.sup..function.) (56)
u
A function V is also conveniently defined in the enlarged space of
variables as
V=W.sup..function. -.THETA..sup.i
(W.sup.i,.tau..sup.i,T.sup.i,.tau..sup..function.,T.sup..function.)=W.sup.
.function. -W.sup.i -I(.tau..sup.i
T.sup.i,.tau..sup..function.,T.sup..function.) (57)
Two mutually equal maxima at the constant W.sup.i and at the constant
W.sup.f are described by the extremal functions
V.sup.i
(W.sup.i,.tau.,T.sup.i,.tau..sup..function.,T.sup..function.)=V.sup.
.function.
(.tau..sup.i,T.sup.i,W.sup..function.,.tau..sup..function.,T.sup.
.function.).tbd.0 (58))
which vanish identically along all optimal paths. These functions can be
written in terms of a wave-function V as follows:
maxV=max{W.sup..function. -W.sup.i
-I(.tau..sup..function.,T.sup..function.,.tau..sup.i,T.sup.i)}=0 (59)
The equation for the integral work function I in the narrow space of the
coordinates (T,.tau.), which does not involve the coordinates W, follows
immediately from the condition V=0. (Development of the backward DP
algorithm and equations for .THETA..sup.i (X.sup.i) are provided later in
the discussion that follows.)
Solution of the HJB equations begins by noting that the system of three
state equations in equation (55) have the state variables x.sub.0 =W,
x.sub.1 =T and x.sub.2 =.tau.. The state equations from equation (55) can
be in a general form as
##EQU47##
For small .DELTA..tau., from equation (60) it follows that
.DELTA.x.sub..beta. =.function..sub..beta.
(x,u).DELTA..tau.+O(.epsilon..sup.2) (61)
In equation (61), the symbol O(.epsilon..sup.2) means second-order and
higher terms. The second-order and higher terms posses the property that
##EQU48##
The optimal final work for the whole path in the interval
[.tau..sup.i,.tau..sup.f ] is the maximum of the criterion
W.sup..function..tbd..THETA..sup.i (x.sup.i +.DELTA.x)=.THETA..sup.i
(W.sup.i +.DELTA.W,T.sup.i +.DELTA.T,.tau..sup.i +.DELTA..tau.) (63)
Expanding equation (63) in a Taylor series yields
##EQU49##
Substituting equation (62) into equation (64) and performing an appropriate
extremization in accordance with Bellman's principle of optimality,
yields, for the variation of the initial point
##EQU50##
After reduction of .THETA..sup.i, and the division of both sides of
equation (65) by .DELTA..tau., the passage to the limit
.DELTA..tau..fwdarw.0 subject to the condition
##EQU51##
yields the backward HJB equation of the optimal control problem. For the
initial point of the extremal path, the backward DP equation is
##EQU52##
The properties of V=W.sup.i -.THETA..sup.i have been used in the second
line of the above equation.
The partial derivative of V with respect to the independent variable .tau.
can remain outside of the bracket of this equation as well. Using the
relationships
##EQU53##
and W=.function..sub.0 =-L (see e.g., equations (47) and (50)), leads to
the extremal work function
##EQU54##
In terms of the integral function of optimal work, I=W.sup..function.
-W.sup.i -V, equation (69) becomes
##EQU55##
As long as the optimal control u is found in terms of the state, time, and
gradient components of the extremal performance function I, the passage
from the quasi-linear HJB equation to the corresponding nonlinear
Hamilton-Jacobi equation is possible.
The maximization of equation (70) with respect to the rate u leads to two
equations, the first of which describes the optimal control u expressed
through the variables T and z=.differential.I/.differential.T as follows:
##EQU56##
and the second is the original equation (70) without the extremization sign
##EQU57##
In the proceeding two equations, the index i is omitted. Using the
momentum-tape variable z=.differential.I/.differential.T, and using
equation (71) written in the form
##EQU58##
leads to the energy-type Hamiltonian of the extremal process as
H=zu(z,T)+.function..sub.0 (z,T) (74)
Using the energy-type Hamiltonian and equation and equation (72) leads to
the Hamilton-Jacobi equation for the integral I as
##EQU59##
Equation (75) differs from the HJB equation in that equation (75) refers to
an external path only, and H is the extremal Hamiltonian. The above
formulas are applied to the concrete a Lagrangian L=-.function..sub.0,
where .function..sub.0 is the intensity of the mechanical work production.
The basic integral W.sub.[T.sub..sup.i .sub.,.tau..sub..sup..function.
.sub.] written in the form of equation (47)
##EQU60##
whose extremal value is the function
I(T.sup.i,.tau..sup.i,T.sup..function.,.tau..sup..function.). The
momentum-like variable (equal to the temperature adjoint) is then
##EQU61##
and
##EQU62##
The Hamilton-Jacobi partial differential equation for the maximum work
problem (engine mode of the system) deals with the initial coordinates,
and has the form
##EQU63##
The variational front-end problem for the maximum work W is equivalent to
the variational fixed-end problem of the minimum entropy production.
Hamilton-Jacobi Approach to Minimum Entropy Production
The specific entropy production is described by the functional from
equation (49) as
##EQU64##
The minimum of the functional can be described by the optimal function
I.sub..sigma. (T.sup.i,.tau..sup.i,T.sup..function.,.tau..sup..function.))
and the Hamilton-Jacobi equation is then
##EQU65##
The two functionals (that of the work and that of the entropy generation)
yield the same extremal.
Considering the Hamilton-Jacobi equation for the engine, from equation (54)
integration along an extremal path leads to a function which describes the
optimal specific work
##EQU66##
where the parameter .xi. is defined in equation (53). The extremal specific
work between two arbitrary states follows in the form
##EQU67##
From equation (83), with T.sup.i =T.sup.w and T.sup.f =T.sup.A, then the
minimal integral of the entropy production is
##EQU68##
The function from equation (83) satisfies the backward Hamilton-Jacobi
equation, which is equation (79), and the function (84) satisfies the
Hamilton-Jacobi equation (81).
The Lyapunnov function .Fourier. for the system of equation (60) can be
defined as
##EQU69##
and
##EQU70##
The relation (86) is the necessary and sufficient conditions for stability
and robustness; of control with minimum entropy in the control object and
the control system.
Suspension Control
In one embodiment, the reduced control system of FIGS. 4A and 4B is applied
to a suspension control system, such as, for example, in an automobile,
truck, tank, motorcycle, etc.
FIG. 8 is a schematic diagram of one half of an automobile suspension
system. In FIG. 8, a right wheel 802 is connected to a right axle 803. A
spring 804 controls the angle of the axle 803 with respect to a body 801.
A left wheel 812 is connected to a left axle 813 and a spring 814 controls
the angle of the axle 813. A torsion bar 820 controls the angle of the
left axle 803 with respect to the right axle 813.
FIG. 9 is a block diagram of a control system for controlling the
automobile suspension system shown in FIG. 8. The block diagram of FIG. 9
is similar to the diagram of FIG. 4B showing an optimal control system and
a reduced control system 980.
The optimal control system 920 includes an optimal sensor system 922 that
provides information I.sub.1 to an input of a control object model 924. An
output of the control object model 924 is provided to a first genetic
analyzer GA1927, and the GA1927 provides an off-line tuning signal to a
first fuzzy neural network FNN1926. An output from the FNN1926, and an
output from the optimal sensor system 922 are provided to a control unit
925. An output from the control unit 925 is provided to a control object
928. The control object 928 is the automobile suspension and body shown in
FIG. 8.
The reduced control system 980 includes a reduced sensor system 982 that
provides reduced sensor information I.sub.2 to an input of a control unit
985. The control unit 985 is configured by a second fuzzy neural network
FNN2986. An output from the control unit 985 is provided to a control
object 988. The control object 988 is also the automobile suspension and
body shown in FIG. 8.
The FNN2986 is tuned by a second genetic analyzer GA2, that, (like the
GA2444 in FIG. 4.) maximizes an information signal I(x,y) and minimizes an
entropy signal S. The information signal I(x,y) is computed from a
difference between the control and sensor signals produced by the optimum
control system 920 and the reduced control system 980. The entropy signal
:S is computed from a mathematical model of the control object 988.
The optimal sensor system 922 includes a pitch angle sensor, a roll angle
sensor, four position sensors, and four angle sensors as described below.
The reduced sensor system 922 includes a single sensor, a vertical
accelerometer placed near the center of the vehicle 801.
A Half Car Coordinate Transformation
1. Description of transformation matrices
1.1 Global reference coordinate x.sub.r, y.sub.r, z.sub.y {r} is assumed to
be at the pivot center P.sub.r of a rear axle.
The following are the transformation matrices to describe the local
coordinates for:
{2} the gravity center of the body.
{7} the gravity center of the suspension.
{10} the gravity center of the arm.
{12} the gravity center of the wheel.
{13} the touch point of the wheel to the road.
{14} the stabilizer linkage point.
1.2 Transformation matrices.
Rotating {r} along y.sub.r with angle .beta. makes a local coordinate
system x.sub.0r, y.sub.0r, z.sub.0r {0r} with a transformation matrix
.sup.r.sub.0r T.
##EQU71##
Transferring {0r} through the vector (a.sub.1, 0, 0) makes a local
coordinate system x.sub.0f, y.sub.0f, z.sub.0f {0f} with a transformation
matrix .sup.0r.sub.0f T.
##EQU72##
The above procedure is repeated to create other local coordinate systems
with the following transformation matrices.
##EQU73##
1.3 Coordinates for the front wheels 802, 812 (index n:i for the left, ii
for the right) are generated as follows.
Transferring {1f} through the vector (0, b.sub.2n, 0) makes local
coordinate system x.sub.3n, y.sub.3n, z.sub.3n {3n} with transformation
matrix .sup.1f.sub.3n T.
##EQU74##
1.4 Some matrices are sub-assembled to make the calculation simpler.
##EQU75##
2. Description of all the parts of the model both in local coordinate
systems and relations to the coordinate{r} or {1f}.
2.1 Description in local coordinate systems.
##EQU76##
2.2 Description in global reference coordinate system {r}.
##EQU77##
where .zeta..sub.n is substituted by,
.zeta..sub.n =-.gamma..sub.n -.theta..sub.n
because of the link mechanism to support wheel at this geometric relation.
2.3 Description of the stabilizer linkage point in local coordinate system
{1f}.
The stabilizer works as a spring in which force is proportional to the
difference of displacement between both arms in a local coordinate system
{1f} fixed to the body.
##EQU78##
3. Kinetic energy, potential energy and dissipative functions for the
<Body>, <Suspension>, <Arm>, <Wheel> and <Stabilizer>.
Kinetic energy and potential energy except by springs are calculated based
on the displacement referred to the inertial global coordinate {r}.
Potential energy by springs and dissipative functions are calculated based
on the movement in each local coordinate.
##EQU79##
where
##EQU80##
and
##EQU81##
and thus
##EQU82##
##EQU83##
##EQU84##
##EQU85##
##EQU86##
##EQU87##
where
##EQU88##
and
##EQU89##
thus
##EQU90##
where
##EQU91##
Substituting m.sub.an with m.sub.wn and e.sub.1n with e.sub.3n in the
equation for arm, yields an equation for the wheel as:
##EQU92##
Therefore the total kinetic energy is:
##EQU93##
where
##EQU94##
Hereafter variables and constants which have index `n` implies implicit or
explicit that they require summation with n=i, ii
Total potential energy is:
##EQU95##
where
m.sub.ba =m.sub.b (.alpha..sub.0 +.alpha..sub.1)
m.sub.sawan =(m.sub.sn +m.sub.an +m.sub.wn).alpha..sub.1n
m.sub.sawbn =(m.sub.sn +m.sub.an +m.sub.wn)b.sub.2n
m.sub.sawcn =m.sub.sn c.sub.1n +(m.sub.an +m.sub.wn)c.sub.2n
.gamma..sub.ii =-.gamma..sub.i (147)
4. Lagrange's Equation
The Lagrangian is written as:
##EQU96##
##EQU97##
##EQU98##
##EQU99##
##EQU100##
##EQU101##
##EQU102##
##EQU103##
##EQU104##
##EQU105##
##EQU106##
##EQU107##
##EQU108##
##EQU109##
##EQU110##
The dissipative function is:
##EQU111##
The constraints are based on geometrical constraints, and the touch point
of the road and the wheel. The geometrical constraint is expressed as
e.sub.2n cos .theta..sub.n =-(z.sub.6n -d.sub.1n) sin .eta..sub.n
e.sub.2n sin .theta..sub.n -(z.sub.6n -d.sub.1n) cos .eta..sub.n =c.sub.1n
-c.sub.2n (168)
The touch point of the road and the wheel is defined as
z.sub.in =z.sub.p.sub..sub.touchpoint n .sub..sup.r
={z.sub.12n cos .alpha.+e.sub.3n sin (.alpha.+.gamma..sub.n
+.theta..sub.n)+c.sub.2n cos (.alpha.+.gamma..sub.n)
+b.sub.2n sin .alpha.} cos .beta.-.alpha..sub.1n sin .beta.
=R.sub.n (t) (169)
where R.sub.n (t) is road input at each wheel.
Differentials are:
.theta..sub.n e.sub.2n sin .theta..sub.n -z.sub.6n sin .eta..sub.n
-.eta..sub.n (z.sub.6n -d.sub.1n) cos .eta..sub.n =0
.theta..sub.n e.sub.2n cos .theta..sub.n -z.sub.6n cos .eta..sub.n
+.eta..sub.n (z.sub.6n -d.sub.1n) sin .eta..sub.n =0
{z.sub.12n cos .alpha.-.alpha.z.sub.12n sin
.alpha.+(.alpha.+.theta..sub.n)e.sub.3n cos (.alpha.+.gamma..sub.n
+.theta..sub.n)
-.alpha.c.sub.2n sin (.alpha.+.gamma..sub.n)+.alpha.b.sub.2n cos .alpha.}
cos .beta.
-.beta.[{z.sub.12n cos .alpha.+e.sub.3n sin (.alpha.+.gamma..sub.n
+.theta..sub.n)
+c.sub.2n cos (.alpha.+.gamma..sub.n)+b.sub.2n sin .alpha.} sin
.beta.+.alpha..sub.1n cos .beta.]-R.sub.n (t)=0 (170)
Since the differentials of these constraints are written as
##EQU112##
then the values a.sub.1nj are obtained as follows.
.alpha..sub.1n1 =0, .alpha..sub.1n2 =0, .alpha..sub.1n3 =-(z.sub.6n
-d.sub.1n) cos .eta..sub.n, .alpha..sub.1n4 =e.sub.2n sin .theta..sub.n,
.alpha..sub.1n5 =-sin .eta..sub.n, .alpha..sub.1n6 =0
.alpha..sub.2n1 =0, .alpha..sub.2n2 =0, .alpha..sub.2n3 =(z.sub.6n
-d.sub.1n) sin .eta..sub.n, .alpha..sub.2n4 =e.sub.2n cos .theta..sub.n,
.alpha..sub.2n5 =-cos .theta..sub.n, .alpha..sub.2n6 =0
.alpha..sub.3n1 =-{z.sub.12n cos .alpha.+e.sub.3n sin
(.alpha.+.gamma..sub.n +.theta..sub.n)+c.sub.2n cos
(.alpha.+.gamma..sub.n)+b.sub.2n sin .alpha.} sin .beta.+.alpha..sub.1n
cos .beta.,
.alpha..sub.3n2 ={-z.sub.12n sin .alpha.+e.sub.3a cos
(.alpha.+.gamma..sub.n +.theta..sub.n)-c.sub.2n sin
(.alpha.+.gamma..sub.n)+b.sub.2n cos .alpha.} cos .beta.,
.alpha..sub.3n3 =0, .alpha..sub.3n4 =e.sub.3n cos (.alpha.+.gamma..sub.n
+.theta..sub.n) cos .eta., .alpha..sub.3n5 =0, .alpha..sub.3n6 =cos
.alpha. cos .eta. (172)
From the above, Lagrange's equation becomes
##EQU113##
where
q.sub.1 =.beta., q.sub.2 =.alpha., q.sub.3i =.eta..sub.i, q.sub.4i
=.theta..sub.i, q.sub.5i =z.sub.6i, q.sub.6i =z.sub.12i
q.sub.3ii =.eta..sub.ii, q.sub.4ii =.theta..sub.ii, q.sub.5ii =z.sub.6ii,
q.sub.6ii =z.sub.12ii (174)
can be obtained as follows:
##EQU114##
##EQU115##
##EQU116##
##EQU117##
##EQU118##
##EQU119##
##EQU120##
##EQU121##
##EQU122##
##EQU123##
##EQU124##
##EQU125##
##EQU126##
##EQU127##
##EQU128##
##EQU129##
##EQU130##
##EQU131##
##EQU132##
##EQU133##
##EQU134##
##EQU135##
##EQU136##
##EQU137##
From the differentiated constraints it follows that:
.theta..sub.n e.sub.2n sin .theta..sub.n -z.sub.6n sin .eta..sub.n
-.eta..sub.n (z.sub.6n -d.sub.1n) cos .eta..sub.n =0
.theta..sub.n e.sub.2n cos .theta..sub.n -z.sub.6n cos .eta..sub.n +.sub.n
(z.sub.6n -d.sub.1n) sin .eta..sub.n =0 (200)
and
{z.sub.12n cos .alpha.-.alpha.z.sub.12n sin
.alpha.+(.alpha.+.theta..sub.n)e.sub.3n cos (.alpha.+.gamma..sub.n
+.theta..sub.n)
-.alpha.c.sub.2n sin (.alpha.+.gamma..sub.n)+.alpha.b.sub.2n cos .alpha.}
cos .beta.
-.beta.[{z.sub.12n cos .alpha.+e.sub.3n sin (.alpha.+.gamma..sub.n
+.theta..sub.n)
+c.sub.2n cos (.alpha.+.gamma..sub.n)+b.sub.2n sin .alpha.} sin
.eta.+.alpha..sub.1n cos .beta.]-R.sub.n (t)=0 (201)
Second order of differentiation of equation (200) yields:
##EQU138##
and
##EQU139##
Supplemental differentiation of equation (199) for the later entropy
production calculation yields:
k.sub.wn z.sub.12n =-c.sub.wn z.sub.12n +.lambda..sub.3n C.sub..alpha.
C.sub..beta. -.alpha..lambda..sub.3n S.sub..alpha. C.sub..beta.
-.beta..lambda..sub.3n C.sub..alpha. S.sub..beta. (206)
therefore
##EQU140##
or from the third equation of constraint:
{z.sub.12n cos .alpha.-z.sub.12n.alpha. cos .alpha.-.alpha.z.sub.12n sin
.alpha.-.alpha.z.sub.12n sin .alpha.-.alpha..sup.2 z.sub.12n cos
.alpha.+(.alpha.+.theta..sub.n)e.sub.3n cos (.alpha.+.gamma..sub.n
+.theta..sub.n)
-(.alpha.+.theta..sub.n).sup.2 e.sub.3n sin (.alpha.+.gamma..sub.n
+.theta..sub.n)-.alpha.c.sub.2n sin
(.alpha.+.gamma..sub.n)-.alpha.c.sub.2n cos
(.alpha.+.gamma..sub.n)+.alpha.b.sub.2n cos .alpha.-.alpha..sup.2 b.sub.2n
sin .alpha.} cos .beta.
-.beta.{z.sub.12n cos .alpha.-.alpha.z.sub.12n sin
.alpha.+(.alpha.+.theta..sub.n)e.sub.3n cos (.alpha.+.gamma..sub.n
+.theta..sub.n)-.alpha.c.sub.2n sin
(.alpha.+.gamma..sub.n)+.alpha.b.sub.2n cos .alpha.} sin .beta.
-.beta.[{z.sub.12n cos .alpha.+e.sub.3n sin (.alpha.+.gamma..sub.n
+.theta..sub.n)+c.sub.2n cos (.alpha.+.gamma..sub.n)+b.sub.2n sin .alpha.}
sin .beta.+.alpha..sub.1n cos .beta.]
-.beta.[{z.sub.12n cos .alpha.-.alpha.z.sub.12n sin
.alpha.+(.alpha.+.theta..sub.n) cos (.alpha.+.gamma..sub.n
+.theta..sub.n)-(.alpha.+.gamma..sub.n)c.sub.2n sin
(.alpha.+.gamma..sub.n)+.alpha.b.sub.2n cos .alpha.} sin .beta.
+.beta.{z.sub.12n cos .alpha.+e.sub.3n sin (.alpha.+.gamma..sub.n
+.theta..sub.n)+c.sub.2n cos (.alpha.+.gamma..sub.n)+b.sub.2n sin .alpha.}
cos .beta.-.beta..alpha..sub.1n sin .beta.]-R.sub.n (t)=0 (208)
##EQU141##
5. Summarization for simulation.
##EQU142##
##EQU143##
In the above expression, the potential energy of the stabilizer is halved
because there are left and right portions of the stabilizer.
##EQU144##
where
##EQU145##
The initial conditions are:
##EQU146##
where l.sub.sn is free length of suspension spring, l.sub.s0n is initial
suspension spring deflection at one g, l.sub.wn is free length of spring
component of wheel and l.sub.w0n is initial wheel spring deflection at one
g.
IV. Equations for Entropy Production
Minimum entropy production (for use in the fitness function of the genetic
algorithm) is expressed as:
##EQU147##
##EQU148##
##EQU149##
##EQU150##
##EQU151##
Other Embodiments
Although the foregoing has been a description and illustration of specific
embodiments of the invention, various modifications and changes can be
made thereto by persons skilled in the art, without departing from the
scope and spirit of the invention as defined by the following claims.
Top